Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails

2609.09134 ↗ · 2026-09-08 · cs.AI

Zhou Yu, Bin Bi, Shiva Kumar Pentyala, Shubham Mehrotra, Sougata Chaudhuri, Shilpa Bhagavath, Zeyuan Chen, Ran Xu, Phil Mui, James Zhu, Sitaram Asur

The finding

Headline claim

When an agent harness has been automatically optimized around a weaker model, fine-tuning that model on a stronger expert's full trajectories degrades task success, while localized on-policy expert correction of the weaker model's own failed rollouts preserves harness fit and adds small gains on top.

The test

Seven objectively verifiable enterprise agentic benchmarks (payroll auditing, budget approval, stock alerting, IoT anomaly detection, browser automation, website management, code refactoring) using GEPA-style harness evolution with gemini-3.1-pro-preview as meta-agent and expert, and LoRA-SFT on qwen3-coder-30b-a3b (with a gemma-4-26b-a4b replication on Webarena), test success averaged over three runs.

Effect size & scope

Harness evolution lifts weak-model mean success from 29.2% to 78.0% (+48.8 points); the expert reaches 93.6% on the evolved harness; full-trajectory imitation then drops the weak model from 78.0% to 63.1% (all seven tasks regress, -4.2 to -29.9 points, mean -14.9), while the identical recipe helps under the baseline harness (29.2% to 35.5%); on-policy correction recovers to 79.7% (+1.7 over the evolved-harness baseline). Gemma replication shows the same regression pattern (55.6% to 41.1% under the evolved harness). Effect holds within these seven tasks and two model families; the correction gain is modest and not yet iterated across co-evolution rounds in the reported results.

Limitations

The authors concede some absolute results are not directly comparable with Yang et al. (2026) due to environment changes; only one replication task for Gemma; the co-evolution loop is proposed but not exhaustively demonstrated; expert-correction gains are small (+1.7) relative to harness gains (+48.8). Unstated: dependence on a closed frontier model (Gemini) as expert and meta-agent, possibly contaminated or shifting; seven bespoke benchmarks with a shared optimization framework may not generalize; no cost accounting of the meta-agent search or correction pipeline; no comparison to RL-based alternatives beyond SFT.

Counterpoint

A skeptic would note this is Salesforce research promoting an internal pipeline, on seven benchmarks from a cited related framework with the authors' own environment tweaks, with tiny model counts (two families, one expert), three-run averages, and an effect whose practical magnitude (avoiding a self-inflicted regression, then +1.7 points) is thin next to the headline harness-evolution gains borrowed from prior work. Prompt and harness-search stochasticity could contribute, and the mechanistic claim (planning-strategy drift) rests on failure-taxonomy bucketing rather than a causal intervention.

Economy implication

Little or nothing yet for macro AI economics: the actionable piece is an engineering caution for enterprises adapting small open-weight models with evolved harnesses, namely that naive SFT on expert traces can destroy value and cheaper small-model agents require careful co-adaptation. It weakly supports the viability of small-model, harness-heavy deployments as a cost lever, but the gains shown are task-specific and not yet evidence of durable enterprise ROI.

Who this matters for

Technical Capability: It identifies a real interaction effect between scaffolding optimization and weight updates, a concrete pitfall for anyone building agents by stacking both levers.

Labor & Knowledge Work: If smaller open-weight models plus automated harness search can hit 78-80% on enterprise tasks, the cost of automating tool-using knowledge work falls further, which matters for which roles get automated first.

Bears on (advisory · papers never write evidence)

[5.2] Open-weight models reach "good enough" parity for most economic use cases.

[B1] Agentic reliability gains will translate into enterprise ROI.

Abstract

Agent harnesses (the system prompt, tool set, execution hooks, and context-management scaffolding around a model) are a critical determinant of agentic task success. Automated harness evolution can enable smaller models to perform well on domain-specific tasks at a fraction of frontier-model cost. Since both the harness and model weights shape behavior, we ask how harness evolution and lightweight fine-tuning should be combined. Across seven enterprise agent tasks, we first evolve a harness with the weaker model, then find that a stronger expert often uses it more effectively, suggesting expert supervision could close the remaining gap. However, training the weaker model on the expert's complete trajectories under the evolved harness backfires: performance regresses on all seven tasks by 4 to 30 points across Qwen3-Coder and Gemma 4, even though the same procedure helps under the unevolved harness. Our analysis shows that imitation transfers knowledge and increases scaffold usage, but disrupts model-harness fit: the weaker model adopts the expert's planning strategy without the competence to execute it and no longer matches the harness evolved around its native planning style. We therefore develop an on-policy expert-correction pipeline, automated by a meta-level MLE agent, that localizes the failing turn in the weaker model's own rollout and asks the expert to rewrite only that turn. This preserves the model's planning style and combines the gains of harness evolution and model adaptation. Our results identify and resolve a source of contention between harness and weight updates, yielding a compatibility-preserving recipe for economical co-evolution on domain-specific enterprise tasks.